FaithSCAN: Model-Driven Single-Pass Hallucination Detection for Faithful Visual Question Answering

TL;DR

FaithSCAN targets faithfulness hallucinations in VQA by harnessing multiple internal uncertainty signals from a single forward pass of a vision-language model. It fuses token-level decoding uncertainty, visual representations, and cross-modal alignment through branch-wise encoders and uncertainty-aware attention, trained with model-driven supervision via Visual-NLI. The approach achieves strong in-distribution performance and competitive out-of-distribution results while offering interpretability through token-level attributions and analysis of internal states. This work demonstrates that hallucination detection can be efficient and robust by treating hallucinations as intrinsic multimodal reasoning phenomena and leveraging internal model signals rather than external verification or repeated sampling.

Abstract

Faithfulness hallucinations in VQA occur when vision-language models produce fluent yet visually ungrounded answers, severely undermining their reliability in safety-critical applications. Existing detection methods mainly fall into two categories: external verification approaches relying on auxiliary models or knowledge bases, and uncertainty-driven approaches using repeated sampling or uncertainty estimates. The former suffer from high computational overhead and are limited by external resource quality, while the latter capture only limited facets of model uncertainty and fail to sufficiently explore the rich internal signals associated with the diverse failure modes. Both paradigms thus have inherent limitations in efficiency, robustness, and detection performance. To address these challenges, we propose FaithSCAN: a lightweight network that detects hallucinations by exploiting rich internal signals of VLMs, including token-level decoding uncertainty, intermediate visual representations, and cross-modal alignment features. These signals are fused via branch-wise evidence encoding and uncertainty-aware attention. We also extend the LLM-as-a-Judge paradigm to VQA hallucination and propose a low-cost strategy to automatically generate model-dependent supervision signals, enabling supervised training without costly human labels while maintaining high detection accuracy. Experiments on multiple VQA benchmarks show that FaithSCAN significantly outperforms existing methods in both effectiveness and efficiency. In-depth analysis shows hallucinations arise from systematic internal state variations in visual perception, cross-modal reasoning, and language decoding. Different internal signals provide complementary diagnostic cues, and hallucination patterns vary across VLM architectures, offering new insights into the underlying causes of multimodal hallucinations.

Figures (9)

• Figure 1: Illustration of three representative VQA scenarios: (a) object existence questions, (b) relationships between two objects within a single image, and (c) reasoning with a false-premise question. Modern VLM-based QA systems perform reliably on simple questions, but often produce unfaithful or hallucinated answers in complex scenarios.
• Figure 2: Overview of FaithSCAN for single-pass hallucination detection in VQA. (a) Workflow of our method: multiple types of internal uncertainty signals are extracted from a frozen VLM in a single forward pass and combined with model-driven supervision to train FaithSCAN. (b) Core components of FaithSCAN, including the construction of supervision signals and the model architecture.
• Figure 3: RQ2: Illustrative examples of the consistency judgment process based on model-driven supervision, showing both positive and negative cases. (a) The sample is drawn from POPE, where the LLM-based judger correctly identifies a contradiction relationship between the model-generated answer and the ground-truth reference. (b) The sample is drawn from VQA v2, where the generated answer is not strictly precise; nevertheless, it is judged as entailment by the LLM judger.
• Figure 4: RQ3: AUROC scores for different uncertainty source combinations used in hallucination detection across four datasets and three base models. Labels indicate uncertainty types: T = text-level (log-likelihood, entropy, embedding), P = patch-level visual, and A = cross-modal alignment.
• Figure 5: RQ3: Ablation study on feature combination, sequence modeling strategy, and fusion methods on the HalLoc-VQA training set. The fusion methods include: Global Avg (average of all token-level features), CNN Emb (CNN modeling only on the embeddings), and CNN All (CNN modeling applied to all features).